Background: Next-generation sequencing (NGS) assays offer efficient high-throughput profiling methods for analyzing circulating tumor DNA (ctDNA) in blood or other liquid biopsy specimens. The ctDNA mutant allele fractions of patients after treatment is often very low and below the detection limit of regular cfDNA assays. Here we report the development and clinical application of a CLIA-certified ultra-sensitive NGS cfDNA assay in detecting the low-frequency ctDNA mutations and monitoring therapy response and drug resistance in patients with metastatic gastroesophageal adenocarcinoma under combined blockade treatments. Methods: PredicineCARE ULTRA cfDNA assay with 100,000X sequencing depth was validated under CLIA settings. Limit of blank (LoB) and limit of detection (LoD) were established by profiling 38 healthy donor plasma samples, 23 contrived human cfDNAs samples from patients with various cancers, and 30 replicates of well-established ctDNA reference standards. In silico down-sampling of FASTQ files was carried out to test the sensitivity of this assay with a lower coverage (50,000X). This assay was then adopted for longitudinal sequencing of 57 serial plasmas from 6 patients with HER2+ gastroesophageal adenocarcinomas (GEC) who underwent combined treatments (capecitabine, oxaliplatin, bevacizumab and trastuzumab) at different time points (baseline, response, and progression). Results: PredicineCARE ULTRA cfDNA assay exhibited exceptional sensitivity in calling single-nucleotide variants and indels, with a 98.0% detection rate at a LoD of 0.075% mutant allele frequency (MAF) for 60 ng cfDNA. This is significantly lower than the LoD (0.25% MAF) of a standard PredicineCARE cfDNA assay. The ultra-sensitive assay also displayed decent analytical specificity (>99.99%) and precision (intra-assay 100%; inter-assay 99.41%). In 6 GEC patients’ longitudinal samples, PredicineCARE ULTRA cfDNA assay detected 36 additional somatic mutations with the MAFs spanning in a range from 0.024% to 0.1% beyond 124 baseline mutations identified using standard PredicineCARE. In two representative cases presented here, all baseline mutations (genes TP53 and PIK3CA etc.) were consistently reported in the follow-up time points with a noticeable MAF reduction (lowest MAFs from 0.046% to 0.024%) during response. Subsequent elevations were observed in more genes, including low-frequency mutations (KRAS and ERBB2) not detected at baseline or only detected in PredicineCARE ULTRA. Conclusions: This study demonstrated the promise and clinical utility of the ultra-sensitive cfDNA molecular profiling at 100,000X sequencing depth in treatment selection, therapy monitoring, and drug resistance studies.